[RFC][4.4] PR rtl-optimization/3507 appalling optimization with sub/cmp

[Paolo Bonzini]

> 	(cse_main): Ask DF for def-use chains.

I'm not sure I like using def-use chains in CSE.  The pass works in 
completely different ways.  For example we already have 
find_comparison_args doing the reverse of def-use chains.

/* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
    operation (EQ, NE, GT, etc.), follow it back through the hash table and
    what values are being compared.

    *PARG1 and *PARG2 are updated to contain the rtx representing the values
    actually being compared.  For example, if *PARG1 was (cc0) and *PARG2
    was (const_int 0), *PARG1 and *PARG2 will be set to the objects that 
were
    compared to produce cc0.

    The return value is the comparison operator and is either the code of
    A or the code corresponding to the inverse of the comparison.  */

Second, and more important, the optimization belongs in fold_rtx, just 
before

   rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
   rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
   new = simplify_relational_operation (code, mode, mode_arg0, op0, op1);

and could use find_comparison_args+lookup to extract the information it 
needs.

Paolo